What does an SDR do, and which parts can software take over?
A sales development representative (SDR) sits at the top of the B2B sales funnel. The role exists to create first conversations: pick companies that could buy, reach the person who owns the problem, and hand a qualified meeting to an account executive or a founder. Some companies split the work between SDRs, who often handle inbound leads, and BDRs, who prospect outbound, but the labels vary from one team to the next.
Most of an SDR's week goes to research and writing, not to talking. That is the part an AI SDR targets. The table shows how the work usually divides between the software and your team.
| SDR task | Handled by an AI SDR? | What a person still does |
|---|---|---|
| Building a list of target accounts | Yes, from a database, a CRM export or live signals | Defines the ideal customer profile and the exclusions |
| Researching each account | Yes, from websites, public posts and news | Spot-checks that the research is correct |
| Finding and verifying contacts | Yes, usually with an email verification step | Decides which roles to target |
| Writing first emails and follow-ups | Yes, drafted per prospect | Sets the offer, the tone and the claims that are allowed |
| Sending and pacing | Yes, within limits per inbox | Owns the sending domains and the limits |
| Sorting replies | Yes, by intent (interested, not now, unsubscribe) | Answers real questions and objections |
| Running the discovery call | No | Qualifies the deal and moves it forward |
How does an AI SDR work, step by step?
You write the brief
You describe who you sell to, what problem you solve, which companies to avoid (customers, open deals, competitors, partners) and what the AI may and may not claim. This brief matters more than any setting: a vague brief produces vague emails. Our ideal customer profile template is a good place to start.
It chooses who to contact
A list-based tool filters a contact database by industry, size and job title. A signal-based tool waits for an event, such as a funding round or a public request for a recommendation, and only then picks the company. Either way, the output is a queue of accounts with a reason attached.
It researches the account
The AI reads the company website, recent public posts and news to find something specific to say. Good tools show their sources, so you can check that the 'recent launch' it mentions really happened.
It finds and checks the contact
It identifies the person most likely to own the problem and checks that the business email address exists before sending. Skipping this step is the fastest way to push up your bounce rate.
It writes the sequence
It drafts a first email and a few follow-ups, each built from the research. The best drafts read like a note from someone who did their homework: one observation, one relevant point, one easy question.
It sends, then listens
Emails go out at a controlled pace from connected inboxes. When someone replies, their sequence stops, the reply is labeled, and either a draft answer is proposed or the thread goes to a person.
It learns from the results
Replies, meetings and unsubscribes feed back into targeting and copy. In practice this loop only works if someone reviews the results each week and changes the brief when a segment stays silent.
What does one lead look like, from signal to reply?
Definitions only go so far. Here is a single lead traced through an AI SDR, with the output at each stage and the check a person should make. The seller, the prospect and the post are invented for illustration; the mechanics are what matter.
Assume you sell a customer onboarding tool to US B2B software companies with 10 to 80 employees. Your brief excludes current customers, open deals and agencies, and it forbids the AI from quoting prices or promising integrations.
| Stage | What the AI SDR produces | What you check |
|---|---|---|
| Signal | A public post by a head of customer success: the team is 35 people, onboarding lives in a spreadsheet, and they ask what other teams use | Is the post recent, and is it a genuine question rather than a vendor fishing for leads? |
| Company match | The author's profile links to a project management SaaS; its website and team page confirm B2B software and roughly 35 people | Is it the right company, and not a customer or an open deal? |
| Score | 82 out of 100: size fits, B2B software fits, explicit request within 48 hours; points off because the budget owner is unclear | Do the written reasons match your own judgment? |
| Contact | The post author, with a business email that passes verification | If verification fails, the lead is dropped, not guessed |
| First email | A short note that names the question they asked and offers one useful idea | Every fact is true, and nothing the brief forbids appears |
| Follow-up | Four days later, a one-line note with a sample onboarding checklist and no new ask | Does it add something, or just bump the thread? |
| Reply | "We are looking at options next month. What does it cost?" Labeled interested, with a drafted answer | You answer the pricing question yourself and propose a call |
The first email it drafted
Subject: your onboarding spreadsheet question
Body: Hi Dana, you asked this week how other SaaS teams run onboarding once a spreadsheet stops scaling. The usual breaking point is two CSMs editing one tracker while nobody can tell which steps a new account has finished. Our tool gives each new account its own checklist that updates as steps are completed. Worth a 15-minute look before you choose something?
- The subject line reuses their words, so it reads as a reply to a conversation they started.
- The first sentence says where the context came from. Hiding that you read a public post costs trust the moment they notice.
- The middle sentence describes the problem in your brief's language, not a guess about Dana's team. Claims about the prospect are where AI drafts go wrong, so the safest drafts make few of them.
- The ask is small. A yes-or-no question about 15 minutes is easier to answer than a request for a demo.
What can an AI SDR do well, and where does it fall short?
Where it is strong
- Research at volume. It can read hundreds of websites and posts in the time a person reads ten, and it does not get bored on account number 200.
- Speed. It can react within hours of a public event, while the moment is still fresh.
- Consistency. Follow-ups go out on schedule, replies are logged, and no lead is forgotten in a browser tab.
- First drafts. It writes a specific first version faster than most people, which you can approve or edit.
Where it falls short
- Strategic judgment. It will not tell you that your offer is wrong for a segment. It keeps writing to that segment until you stop it.
- Facts it cannot verify. With thin inputs, a model can state something plausible and false about a prospect or about your product. Restrict the claims it may make and check samples.
- Complex accounts. Large deals with several stakeholders, procurement and internal politics still need a person who can read the room.
- Live conversation. It can answer simple questions by email, but objections, pricing talks and discovery calls belong to your team.
- Fixing a bad list. If the accounts have no reason to care, faster and better-written emails only reach more uninterested people.
What types of AI SDR are there?
Products sold as AI SDRs differ on three questions: where the leads come from, how much the software does without approval, and whether it works outbound or inbound. Knowing which combination you need narrows the field quickly.
List-based or signal-based
A list-based AI SDR starts from a contact database and filters (industry, headcount, title, location). It gives you volume and predictability, but every company in the segment gets similar treatment, whether or not it has a need right now.
A signal-based AI SDR starts from an event: a company launches a product, raises money, hires for a role, or someone posts that they are unhappy with their current tool. Volume is lower and depends on how often your market produces those events, but each message has a reason to exist. Our guide to buying signals with examples lists the events worth watching, and signal-based outbound explains the approach in full.
Copilot or autonomous
A copilot prepares the work and a rep approves it: it suggests accounts, drafts emails and flags promising leads, but a person presses send. Amplemarket's Duo is an example built for sales teams that want an assistant rather than a replacement.
An autonomous AI SDR runs the sequence end to end within rules you set. Vendors such as 11x describe their products as digital workers. Many tools, Startories included, let you start in approval mode and switch a funnel to autonomous once its emails are consistently good.
Outbound or inbound
Most AI SDRs prospect outbound. Some work on inbound demand instead: they talk to visitors on your website, qualify them and book meetings. Qualified's Piper is a well-known example. If most of your pipeline already comes from your website, an inbound agent may matter more to you than outbound prospecting.
Reach companies with a reason to buy this week
Startories finds the buying signal, verifies the decision-maker and runs the outreach until they book a call.
Which AI SDR terms should you know?
Vendor pages use the same dozen terms, not always with the same meaning. These are the definitions this guide uses.
- Ideal customer profile (ICP). The written description of the companies you serve best: industry, size, situation, the problem they have, and the companies you exclude. The AI SDR uses it as its filter, so a loose ICP produces loose targeting.
- Buying signal. A public event suggesting a company may need what you sell now: a funding round, a new hire, a launch, a post asking for a recommendation. Intent data is the broader term for such signals collected at scale.
- Enrichment. Filling in what a lead is missing (company size, location, the decision-maker and their role) from other sources.
- Email verification. Checking that an address exists and accepts mail before anything is sent to it. It keeps your bounce rate low, which mailbox providers watch closely.
- Warm-up. Raising the daily volume of a new inbox step by step, so providers see a normal sending pattern before any cold email goes out.
- Reply classification. Labeling each answer by intent (interested, not now, out of office, unsubscribe) so the right next step happens quickly.
- Approval mode. A person approves each email or reply before it is sent. It is the sensible default for any new campaign, whatever tool you use.
- Held meeting. A meeting that took place with the right person. It is a sturdier success measure than meetings booked, since a booking that never happens has no value.
How do you evaluate an AI SDR before signing?
Demos are built on the vendor's best examples. Ask to see the product work on your market instead, and judge it on its output rather than its interface.
Ask for real leads in your niche
Give the vendor your ideal customer profile and ask for 20 to 30 sample leads with the reason each one was chosen. If the reasons are generic ("matches your industry"), the emails will be generic too.
Fact-check the drafts
Read ten draft emails and verify every claim about the prospect. One invented detail in ten is enough to embarrass you once the tool sends thousands.
Look at the sending setup
Ask which domains and inboxes send the emails, how many emails each inbox sends per day, what happens when bounces rise, and whether sequences stop on reply. Never let a tool send cold email from your main company domain.
Check the controls
Can you approve emails before they go out? Can you exclude customers, open deals and competitors? Can you pause one campaign without stopping everything else?
Read the reporting
Sends and opens say little. Look for replies by type, meetings booked and meetings held, broken down by campaign or segment.
Read the contract
Check the minimum term, how usage is counted, what happens to the domains and the data if you leave, and how opt-outs and deletion requests are handled.
What are the risks for deliverability and your brand?
The main risk of an AI SDR is not one bad email. It is a large number of mediocre emails sent quickly from a domain you care about. Mailbox providers watch how recipients react: Google's email sender guidelines ask senders to authenticate their domains, keep reported spam rates low and make unsubscribing easy, and mail from senders that ignore them can be filtered or rejected. Once a domain's reputation drops, your everyday business email can suffer too.
The brand risk is quieter. An email that gets a fact wrong, flatters in an obviously automated way, or pitches a current customer as if they were a stranger tells the reader how you work. Buyers talk to each other, and screenshots travel.
Then there is the law. In the US, commercial email must follow the CAN-SPAM Act, which requires accurate sender information, an honest subject line, a postal address and a working opt-out (FTC compliance guide). If you write to people in the EU or the UK, data protection rules cover business contacts as well, so check the basis on which you process their data.
- Send from separate, warmed-up domains, never from your main one.
- Cap the daily volume per inbox and stop every sequence on reply.
- Verify each address and pause sending when bounces rise.
- Keep one suppression list that covers every campaign, not just the current one.
- Approve emails by hand until a campaign has proved itself.
When should you not use an AI SDR?
An AI SDR multiplies what you already have. If the offer, the list or the message does not work, it multiplies that too. Hold off in these cases:
- You have not sold the product yourself yet. Until a founder has closed the first deals by hand, nobody knows which pain to lead with. Start with founder-led sales for your first B2B customers.
- Your market is a few hundred named accounts. When each account deserves its own plan, account-based work by a person beats automated sequences.
- Your deals are very small. If a customer is worth less than a few hundred dollars, the cost of a meeting, human or automated, is hard to recover. Run the numbers with our SDR cost and cost-per-meeting guide.
- You sell to consumers. Cold outreach to individuals raises legal and reputation problems that B2B prospecting does not.
- Every message needs legal review. In some regulated fields, each claim must be approved. An AI SDR can still research and draft, but it should not send on its own.
- Nobody can take the meetings. Booked calls that wait a week for an answer go cold. Make sure someone owns the follow-through.
How to start with an AI SDR
Pick one segment, one offer and one source of leads. Write the brief, review the first 50 emails yourself, and judge the result on meetings held after a few weeks rather than on opens. If it works, add a second segment. If it does not, change the targeting before you rewrite the copy.
If you want to test the signal-based approach, Startories detects public buying signals, checks each company against your profile, verifies the decision-maker and drafts emails you can approve before they go out. Plans start at $99 a month, and the first project on Starter starts with a 3-day full-access trial for $1: see pricing. To weigh software against a hire, read AI SDR vs human SDR.
Frequently asked questions
Is an AI SDR the same as a sales chatbot?
No. A chatbot answers people who come to your website. An AI SDR goes out and starts conversations: it picks accounts, writes to them and follows up. Some inbound AI SDRs combine both, talking to website visitors and booking meetings with them.
Will an AI SDR replace my sales team?
It replaces much of the research and writing that SDRs do, not the people who run discovery calls, negotiate and close. Most small teams use one so that a founder or account executive spends time on qualified conversations instead of list building.
How long does it take to set up an AI SDR?
The software can run within days, but the brief takes longer: the ideal customer profile, the offer, the exclusions and the sending domains. New domains and inboxes also need a warm-up period before they can send at their normal volume.
Do AI SDRs write emails that sound automated?
They can, especially when the only inputs are a job title and an industry. Emails improve a lot when the AI has a specific, recent fact to work from, such as a launch or a public request, and when a person reviews the early drafts.
Is it legal to send cold emails with an AI SDR?
In the US, B2B cold email is allowed if it follows CAN-SPAM: truthful sender information and subject lines, a postal address and a working opt-out. Other countries have stricter rules, so check the law where your prospects are based.
How much does an AI SDR cost?
Pricing models vary: some vendors publish monthly plans, others quote a contract after a demo, and engagement platforms charge per user. Startories plans start at $99 a month. Compare any option on cost per held meeting, including the hours you spend reviewing its work.